imports¶
In [1]:
import os
import sys
import pickle
import numpy as np
import matplotlib as mpl
import cv2
import itertools
import six
import pickle
import utils
import tensorflow as tf
from tensorflow.contrib.layers import flatten, l2_regularizer, conv2d, max_pool2d, avg_pool2d, dropout, fully_connected
from tensorflow.python.client import device_lib
from sklearn.utils import shuffle
import pandas as pd
def get_available_gpus():
local_device_protos = device_lib.list_local_devices()
return [x.name for x in local_device_protos if x.device_type == 'GPU']
font = {'weight' : 'bold',
'size' : 22}
mpl.rc('font', **font)
mpl.rcParams['text.color'] = 'g'
get_available_gpus()
override = False
Step 0: Load The Data¶
In [2]:
training_file = '../traffic-signs-data/train.p'
validation_file= '../traffic-signs-data/valid.p'
testing_file = '../traffic-signs-data/test.p'
with open(training_file, mode='rb') as f:
train = pickle.load(f)
with open(validation_file, mode='rb') as f:
valid = pickle.load(f)
with open(testing_file, mode='rb') as f:
test = pickle.load(f)
X_train, y_train = train['features'], train['labels']
X_valid, y_valid = valid['features'], valid['labels']
X_test, y_test = test['features'], test['labels']
postprocessed = False
file_list = [os.path.join('web_images', name) for name in os.listdir("web_images/")]
image_list = [f for f in file_list if os.path.splitext(f)[-1] in ['.jpg', '.png', '.jpeg']]
y_web = np.array([int(os.path.split(name)[-1].split("_")[0]) for name in image_list])
X_web_orig = [cv2.cvtColor(cv2.imread(name), cv2.COLOR_RGB2BGR) for name in image_list]
X_web_resized = list(map(lambda image: cv2.resize(image, (32,32)), X_web_orig))
Step 1: Dataset Summary & Exploration¶
The pickled data is a dictionary with 4 key/value pairs:
'features'is a 4D array containing raw pixel data of the traffic sign images, (num examples, width, height, channels).'labels'is a 1D array containing the label/class id of the traffic sign. The filesignnames.csvcontains id -> name mappings for each id.'sizes'is a list containing tuples, (width, height) representing the original width and height the image.'coords'is a list containing tuples, (x1, y1, x2, y2) representing coordinates of a bounding box around the sign in the image. THESE COORDINATES ASSUMING THE ORIGINAL IMAGE. THE PICKLED DATA CONTAINS RESIZED VERSIONS (32 by 32) OF THESE IMAGES
Provide a Basic Summary of the Data Set Using Python, Numpy and/or Pandas¶
In [3]:
### Replace each question mark with the appropriate value.
### Use python, pandas or numpy methods rather than hard coding the results
# TODO: Number of training examples
n_train = y_train.shape[0]
# TODO: Number of validation examples
n_valid = y_valid.shape[0]
# TODO: Number of testing examples.
n_test = y_test.shape[0]
# TODO: What's the shape of an traffic sign image?
image_shape = X_train[0].shape
# TODO: How many unique classes/labels there are in the dataset.
n_classes = len(set(y_train))
print("Number of training examples =", n_train)
print("Number of validation examples =", n_valid)
print("Number of testing examples =", n_test)
print("Image data shape =", image_shape)
print("Number of classes =", n_classes)
Include an exploratory visualization of the dataset¶
Visualize the German Traffic Signs Dataset using the pickled file(s)
- show all classes
- show samples from selected classes to visualize in-class variance
Data exploration visualization code goes here.¶
In [4]:
import matplotlib.pyplot as plt
# Visualizations will be shown in the notebook.
%matplotlib inline
#X_train[0]
# def setup_figure(m_pics, w, h, n=10):
# return plt.subplots(1, m_pics, squeeze=True, sharex=True, sharey=True, figsize=(m_pics * n, h / w * n))
#utils.setup_figure(1, )
#plt.imshow(X_train)
assert len(set([x.shape for x in X_train])) == 1
signnames = pd.read_csv("signnames.csv", index_col='ClassId')
h, w = X_train[0].shape[:2]
n = 3
nx = 8
ny = int(np.ceil(len(set(y_train)) / nx))
fig, ax = utils.setup_figure(x_pics=nx, y_pics=ny, w=w, h=h, n=n)
visited = set()
for isample in sorted((set(y_train))):
ix = isample % nx
iy = int((isample - ix) / nx)
ind = np.where(y_train == isample)[0]
ind = ind[np.random.randint(len(ind))]
x = X_train[ind]
ax[iy][ix].imshow(x)
ax[iy][ix].set_title("{}:{}".format(isample, signnames.loc[isample, "SignName"]), fontsize=12, color='g')
plt.savefig('output_images/1_raw_samples.png', bbox_inches='tight')
flipleft_to_right = [11, 12, 13, 15, 17, 18, 22, 26, 30, 35]
flip_up_down = [1, 5, 12, 15, 17]
flip_ud_lr = [15, 17, 32, 40]
pairs_flip_lr = {19:20, 20:19, 33: 34, 34:33, 37:36, 36:37, 39:38, 38:39}
In [5]:
assert len(set([x.shape for x in X_train])) == 1
figname = 'output_images/1_raw_images.png'
if os.path.exists(figname) and False:
image = cv2.imread(figname)
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
plt.figure(figsize=(20,20))
plt.imshow(image)
with open('cache/display_images.pkl', 'rb') as fd:
display_images = pickle.load(fd)
else:
h, w = X_train[0].shape[:2]
# fig, ax = utils.setup_figure(x_pics=5, y_pics=2, w=w, h=h)
n = 3
nx = 10
ny = 5
# fit, ax = plt.subplots(ny, nx, squeeze=True, sharex=True, sharey=True, figsize=(nx * n, ny * n))#h / w * n))
fig, ax = utils.setup_figure(x_pics=nx, y_pics=ny, w=w, h=h, n=3)
display_images = []
visited = set()
for iy in range(ny):
j = np.random.randint(ny)
while j in visited:
j = np.random.randint(ny)
visited.add(j)
image_list_iy = []
ind = np.where(y_train == j)[0]
for i in range(nx):
ix = np.random.randint(len(ind))
ax[iy][i].imshow(X_train[ind[ix]])
ax[iy][i].set_title(signnames.loc[j, "SignName"], fontsize=12, color='g')
image_list_iy.append(ind[ix])
display_images.append(image_list_iy)
plt.savefig('output_images/1_raw_images.png', bbox_inches='tight')
with open('cache/display_images.pkl', 'wb') as fd:
pickle.dump(display_images, fd)
In [6]:
# Train
sign_count_train = signnames.copy()
sign_count_train.loc[:, "count"] = pd.Series(y_train).value_counts().sort_index()
sign_count_train = sign_count_train.sort_values(by='count', ascending=True)
sign_count_train.plot.barh(x='SignName', y='count', figsize=(20,30), title='Train')
ax = plt.gca()
ax.tick_params(axis='y', colors='g')
ax.tick_params(axis='x', colors='g')
plt.savefig('output_images/1_train_ordered_hist.png', bbox_inches='tight')
# Validation
sign_count_valid = signnames.copy()
sign_count_valid.loc[:, "count"] = pd.Series(y_valid).value_counts().sort_index()
sign_count_valid = sign_count_valid.sort_values(by='count', ascending=True)
sign_count_valid.plot.barh(x='SignName', y='count', figsize=(20,30), title='Validation')
ax = plt.gca()
ax.tick_params(axis='y', colors='g')
ax.tick_params(axis='x', colors='g')
plt.savefig('output_images/1_valid_ordered_hist.png', bbox_inches='tight')
# Test
sign_count_test = signnames.copy()
sign_count_test.loc[:, "count"] = pd.Series(y_test).value_counts().sort_index()
sign_count_test = sign_count_test.sort_values(by='count', ascending=True)
sign_count_test.plot.barh(x='SignName', y='count', figsize=(20,30), title='Train')
ax = plt.gca()
ax.tick_params(axis='y', colors='g')
ax.tick_params(axis='x', colors='g')
plt.savefig('output_images/1_test_ordered_hist.png', bbox_inches='tight')
In [7]:
# distribution
figname = 'output_images/1_classes_distribution.png'
if os.path.exists(figname):
image = cv2.imread(figname)
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
plt.figure(figsize=(20,20))
plt.imshow(image)
else:
fig, ax = plt.subplots(1, 3, squeeze=True, sharex=True, sharey=True, figsize=(30, 10))
ax[0].hist(y_train, bins = len(set(y_train)))
ax[0].set_title("Train class distribution", color='g')
ax[1].hist(y_valid, bins = len(set(y_train)), color='r')
ax[1].set_title("Validation class distribution", color='g')
ax[2].hist(y_test, bins = len(set(y_train)), color='g')
ax[2].set_title("Test class distribution", color='g')
plt.savefig(figname, bbox_inches='tight')
Step 2: Design and Test a Model Architecture¶
Design and implement a deep learning model that learns to recognize traffic signs.
The model is trained and tested on the German Traffic Sign Dataset.
The LeNet-5 implementation shown in the classroom was the starting point. It was then re-implemented using tensorflow's contrib layers, since it is more concice and readable.
Preprocessing
- LAB space White balance
- YUV space histogram equalization
- RGB Normalization
output 32x32x3
Augmentation
- Nine random shifts in the [-1, 0, 1] x & y directions
Neural network model
- A deep convolutional multilayer
- An inception multilayer
- Full connected multilayer
- layers' concatenated output from both layers as input to the FC multilayer.
- Layers sizes are a linear of the output size and thus should be adaptive to the classification problem size.
- Regularizations
- Dropout
- L2 weight regulizer
Here is an example of a published baseline model on this problem. It's not required to be familiar with the approach used in the paper but, it's good practice to try to read papers like these.
Pre-process the Data Set (normalization, grayscale, etc.)¶
Minimally, the image data should be normalized so that the data has mean zero and equal variance. For image data, (pixel - 128)/ 128 is a quick way to approximately normalize the data and can be used in this project.
Other pre-processing steps are optional. You can try different techniques to see if it improves performance.
Use the code cell (or multiple code cells, if necessary) to implement the first step of your project.
In [8]:
### Preprocess the data here. It is required to normalize the data. Other preprocessing steps could include
### converting to grayscale, etc.
### Feel free to use as many code cells as needed.
Preprocessing class¶
In [9]:
flip_lr = [11, 12, 13, 15, 17, 18, 22, 26, 30, 35]
flip_ud = [1, 5, 12, 15, 17]
flip_ud_and_lr = [32, 40]
pairs_flip_lr = {19:20, 20:19, 33: 34, 34:33, 37:36, 36:37, 39:38, 38:39}
class PreprocessingPipeline(object):
def __init__(self, augment=True):
self.augment = augment
self.flip_lr = flip_lr
self.flip_ud = flip_ud
self.flip_ud_and_lr = flip_ud_and_lr
self.pairs_flip_lr = pairs_flip_lr
@staticmethod
def whiteBalance(image):
# https://pippin.gimp.org/image-processing/chapter-automaticadjustments.html
lab_image = cv2.cvtColor(image, cv2.COLOR_RGB2LAB)
avg_a = np.average(lab_image[:, :, 1])
avg_b = np.average(lab_image[:, :, 2])
L = lab_image[:,:,0] / 255.
lab_image[:, :, 1] = lab_image[:, :, 1] - (np.average(lab_image[:, :, 1]) - 128) * L * 1.1
lab_image[:, :, 2] = lab_image[:, :, 2] - (np.average(lab_image[:, :, 2]) - 128) * L * 1.1
image_out = cv2.cvtColor(lab_image, cv2.COLOR_LAB2RGB)
image[:] = image_out
@staticmethod
def equalizeHist(gray):
# image_yuv = cv2.cvtColor(image, cv2.COLOR_BGR2YUV)
# image_yuv[:, :, 0] = cv2.equalizeHist(image_yuv[:, :, 0])
# image[:] = cv2.cvtColor(image_yuv, cv2.COLOR_YUV2RGB)
gray[:] = cv2.equalizeHist(gray)
@staticmethod
def equalizeRGBHist(image):
image_yuv = cv2.cvtColor(image, cv2.COLOR_RGB2YUV)
image_yuv[:, :, 0] = cv2.equalizeHist(image_yuv[:, :, 0])
return cv2.cvtColor(image_yuv, cv2.COLOR_YUV2RGB)
@staticmethod
def rgb2gray(image):
return cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
@staticmethod
def normalize(gray):
#gray[:] = cv2.normalize(src=gray, dst=None, alpha=0, beta=1, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8UC1)
gray = (gray - gray.min()) / (gray.max() - gray.min())
@staticmethod
def normalizeRGB(image):
#gray[:] = cv2.normalize(src=gray, dst=None, alpha=0, beta=1, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8UC1)
# image[:] = 0.1 + (0.9 - 0.1) * (image - image.min()) / (image.max() - image.min())
r, g, b = cv2.split(image)
r = cv2.normalize(r, None, alpha=0.1, beta=0.9, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
g = cv2.normalize(g, None, alpha=0.1, beta=0.9, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
b = cv2.normalize(b, None, alpha=0.1, beta=0.9, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
return cv2.merge([r, g, b])
@staticmethod
def fliplr(image):
return cv2.flip(image, 1)
@staticmethod
def flipud(image):
return cv2.flip(image, 0)
@staticmethod
def shift(image, tx, ty):
M = np.float32([[1, 0, tx], [0, 1, ty]])
return cv2.warpAffine(image, M, (image.shape[1], image.shape[0]), flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REFLECT_101)
@staticmethod
def rotate_around_y(image):
rows, cols = image.shape[:2]
w = 0.2
nleft = w * np.random.rand(1) * cols
nright = ((1-w) + w * np.random.rand(1)) * cols
pts1 = np.float32([[0,0],[cols, 0],[0, rows],[cols, rows]])
pts2 = np.float32([[nleft, 0],[nright, 0],[nleft, rows],[nright, rows]])
M = cv2.getPerspectiveTransform(pts1, pts2)
return cv2.warpPerspective(image, M, (cols, cols), flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REPLICATE)
def run(self, image, image_class=None):
_image = image.copy()
self.whiteBalance(_image)
_image = self.equalizeRGBHist(_image)
_image = self.normalizeRGB(_image)
# gray = self.rgb2gray(image_)
# self.normalize(gray)
# self.equalizeHist(gray)
# out = gray
out = _image
if self.augment and image_class is not None:
shift_only_orig = True
augments = []
## shifts
if shift_only_orig:
shift = [-1, 0, 1]
for ys, xs in itertools.product(shift, shift):
# if np.random.rand(1) > 0.5:
# continue
yshift = ys * np.random.rand(1) * 0.1 * image.shape[0]
xshift = xs * np.random.rand(1) * 0.1 * image.shape[1]
augments.append((image_class, self.shift(out, xshift, yshift), "shift %f %f" % (xshift, yshift)))
# can do the following for every one of the shifted ...
## flips
if image_class in self.flip_lr:
augments.append((image_class, self.fliplr(out), "fliplr"))
if image_class in self.flip_ud:
augments.append((image_class, self.flipud(out), "flipud"))
if image_class in self.flip_ud_and_lr:
augments.append((image_class, self.fliplr(self.flipud(out)), "fliplr+flipud"))
if image_class in self.pairs_flip_lr:
augments.append((self.pairs_flip_lr[image_class], self.fliplr(out), "fliplr + class change"))
# run ~ 3 rotation about y
for i in range(5):
augments.append((image_class, self.rotate_around_y(out), "y_rotated"))
## shifts
if False:
shift_augments = []
shift = [-1, 0, 1]
for (image_class, aug_image, name) in augments:
for ys, xs in itertools.product(shift, shift):
yshift = ys * np.random.rand(1) * 0.1 * aug_image.shape[0]
xshift = xs * np.random.rand(1) * 0.1 * aug_image.shape[1]
shift_augments.append((image_class, self.shift(aug_image, xshift, yshift), name + " + shift %f %f" % (xshift, yshift)))
for ys, xs in itertools.product(shift, shift):
yshift = ys * np.random.rand(1) * 0.1 * image.shape[0]
xshift = xs * np.random.rand(1) * 0.1 * image.shape[1]
shift_augments.append((image_class, self.shift(out, xshift, yshift), "shift %f %f" % (xshift, yshift)))
augments += shift_augments
# return gray, augments
return _image, augments
# return gray
return _image
#index = np.random.randint(len(X_train))
#index = np.where(y_train == 17)[0]
index = np.where(y_train == 17)[0]
index = index[np.random.randint(len(index))]
image = X_train[index]
image_class = y_train[index]
preprocessor = PreprocessingPipeline()
out, augments = preprocessor.run(image, image_class)
ny = int(np.ceil(len(augments) / 2 + 1))
nx = 4
fig, ax = utils.setup_figure(x_pics=nx, y_pics=ny, w=w, h=h, n=5)
ax[0][0].imshow(image)
ax[0][0].set_title("original", color='g', fontsize=20)
ax[0][1].imshow(out, cmap='gray')
ax[0][1].set_title("postprocessed", color='g', fontsize=20)
for i in range(0, len(augments)):
ix = i % (nx - 2)
iy = 1 + int((i - ix) / (nx - 2))
ax[iy][ix].imshow(augments[i][1], cmap='gray')
ax[iy][ix].set_title(augments[i][2], color='g', fontsize=20)
# 2nd example
index = np.where(y_train == 1)[0]
index = index[np.random.randint(len(index))]
image = X_train[index]
image_class = y_train[index]
out, augments = preprocessor.run(image, image_class)
ax[0][2].imshow(image)
ax[0][2].set_title("original", color='g', fontsize=20)
ax[0][3].imshow(out, cmap='gray')
ax[0][3].set_title("postprocessed", color='g', fontsize=20)
for i in range(0, len(augments)):
ix = i % (nx - 2)
iy = 1 + int((i - ix) / (nx - 2))
ax[iy][ix + 2].imshow(augments[i][1], cmap='gray')
ax[iy][ix + 2].set_title(augments[i][2], color='g', fontsize=20)
# index = np.random.randint(len(X_train))
# index = index[np.random.randint(len(index))]
# flipleft_to_right = [11, 12, 13, 15, 17, 18, 22, 26, 30, 35]
# flip_up_down = [5, 12, 15, 17]
# flip_ud_lr = [32, 40]
# pairs_flip_lr = {19:20, 20:19, 33: 34, 34:33, 37:36, 36:37, 39:38, 38:39}
figname = "output_images/2_rgb_and_postprocessing.png"
plt.savefig(figname)
In [10]:
figname = "output_images/2_postprocessing.png"
preprocessor = PreprocessingPipeline()
# distribution
#fig, ax = utils.setup_figure(x_pics=3, y_pics=1, w=w, h=h, n=3)
if os.path.exists(figname) and False:
image = cv2.imread(figname)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
plt.figure(figsize=(20,20))
plt.imshow(image, cmap='gray')
else:
fig, ax = utils.setup_figure(x_pics=len(display_images[0]), y_pics=len(display_images), w=w, h=h, n=3)
for iy, disp in enumerate(display_images):
for ix, index in enumerate(disp):
out = preprocessor.run(X_train[index])
#ax[iy][ix].imshow(out, cmap='gray')
ax[iy][ix].imshow(out)
else:
plt.savefig(figname)
preprocess dataset¶
In [11]:
augment = True
preprocessor = PreprocessingPipeline(augment=augment)
cached_x = "cache/X_postprocessed.pkl"
if not postprocessed and not os.path.exists(cached_x):
processed = list(map(preprocessor.run, X_train, y_train))
if augment:
x_train = [p[0] for p in processed]
augments = [p[1] for p in processed]
x_train_aug = []
y_train_aug = []
for i, examples in enumerate(augments):
for example in examples:
y_train_aug.append(example[0])
x_train_aug.append(example[1])
#X_train = np.expand_dims(np.concatenate([x_train, x_train_aug]), axis=len(X_valid.shape))
X_train = np.concatenate([x_train, x_train_aug])
y_train = np.concatenate([y_train, y_train_aug])
else:
#X_train = np.expand_dims(np.array(processed), axis=len(X_valid.shape))
X_train = processed
# do not augment validation & test
# X_valid = np.expand_dims(list(map(preprocessor.run, X_valid)), axis=len(X_valid.shape))
# X_test = np.expand_dims(list(map(preprocessor.run, X_test)), axis=len(X_valid.shape))
X_valid = np.array(list(map(preprocessor.run, X_valid))) #, axis=len(X_valid.shape))
X_test = np.array(list(map(preprocessor.run, X_test))) #, axis=len(X_valid.shape))
X_web = np.array(list(map(preprocessor.run, X_web_resized)))
with open(cached_x, "wb") as fd:
pickle.dump((X_train, y_train, X_valid, X_test, X_web), fd, protocol=4)
postprocessed = True
else:
with open(cached_x, "rb") as fd:
(X_train, y_train, X_valid, X_test, X_web) = pickle.load(fd)
In [12]:
preprocessor = PreprocessingPipeline()
X_web = np.array(list(map(preprocessor.run, X_web_resized)))
Model Architecture¶
In [13]:
### Define your architecture here.
### Feel free to use as many code cells as needed.
In [14]:
def RawTFLeNet(x, out_shape):
# Arguments used for tf.truncated_normal, randomly defines variables for the weights and biases for each layer
mu = 0
sigma = 0.1
# TODO: Layer 1: Convolutional. Input = 32x32x1. Output = 28x28x6.
W1 = tf.Variable(tf.truncated_normal((5, 5, 3, 6), mean=mu, stddev=sigma)) # (height, width, input_depth, output_depth)
b1 = tf.Variable(tf.zeros(6))
layer1 = tf.nn.conv2d(x, W1, strides=[1,1,1,1], padding='VALID') + b1
# TODO: Activation.
layer1 = tf.nn.relu(layer1)
# TODO: Pooling. Input = 28x28x6. Output = 14x14x6.
# wout = (win -f)/s +1 = (28-2)/2+1 = 14
layer1 = tf.nn.max_pool(layer1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID') # ksize (batch, height, width, depth)
# TODO: Layer 2: Convolutional. Output = 10x10x16.
# 10 = (14-F)/1 + 1 -> F = 14+1-10 = 5
W2 = tf.Variable(tf.truncated_normal((5, 5, 6, 16), mean=mu, stddev=sigma)) # (height, width, input_depth, output_depth)
b2 = tf.Variable(tf.zeros(16))
layer2 = tf.nn.conv2d(layer1, W2, strides=[1,1,1,1], padding='VALID') + b2
# TODO: Activation.
layer2 = tf.nn.relu(layer2)
# TODO: Pooling. Input = 10x10x16. Output = 5x5x16.
layer2 = tf.nn.max_pool(layer2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID') # ksize (batch, height, width, depth)
# TODO: Flatten. Input = 5x5x16. Output = 400.
fc0 = flatten(layer2)
# TODO: Layer 3: Fully Connected. Input = 400. Output = 120.
W3 = tf.Variable(tf.truncated_normal(shape=(400, 120), mean=mu, stddev=sigma))
b3 = tf.Variable(tf.zeros(120))
layer3 = tf.add(tf.matmul(fc0, W3), b3)
# TODO: Activation.
layer3 = tf.nn.relu(layer3)
# TODO: Layer 4: Fully Connected. Input = 120. Output = 84.
W4 = tf.Variable(tf.truncated_normal(shape=(120, 84), mean=mu, stddev=sigma))
b4 = tf.Variable(tf.zeros(84))
layer4 = tf.add(tf.matmul(layer3, W4), b4)
# TODO: Activation.
layer4 = tf.nn.relu(layer4)
# TODO: Layer 5: Fully Connected. Input = 84. Output = out_shape.
W5 = tf.Variable(tf.truncated_normal(shape=(84, out_shape), mean=mu, stddev=sigma))
b5 = tf.Variable(tf.zeros(out_shape))
logits = tf.add(tf.matmul(layer4, W5), b5)
return logits
Writing LeNet more in a more compact manner¶
In [15]:
def LeNet(x, out_shape):
layer1 = conv2d(x, num_outputs=6, kernel_size=(5,5), padding="VALID", activation_fn=tf.nn.relu)
layer1 = max_pool2d(layer1, (2,2), padding="VALID")
# TODO: Layer 2: Convolutional. Output = 10x10x16.
layer2 = conv2d(layer1, num_outputs=16, kernel_size=(5,5), padding="VALID", activation_fn=tf.nn.relu)
layer2 = max_pool2d(layer2, (2,2), padding="VALID")
# TODO: Flatten. Input = 5x5x16. Output = 400.
fc0 = flatten(layer2)
# TODO: Layer 3: Fully Connected. Input = 400. Output = 120.
layer3 = fully_connected(fc0, num_outputs=120, activation_fn=tf.nn.relu)
# TODO: Layer 4: Fully Connected. Input = 120. Output = 84.
layer4 = fully_connected(layer3, num_outputs=84, activation_fn=tf.nn.relu)
# TODO: Layer 5: Fully Connected. Input = 84. Output = out_shape.
logits = fully_connected(layer4, num_outputs=out_shape, activation_fn=None)
return logits
In [16]:
def network(x, out_shape, keep_prob, train):
if isinstance(train, bool):
train = tf.constant(train)
# Layer 1: Deep convolutional
_layer1_0 = conv2d(x, num_outputs=16, kernel_size=(3, 3), padding="VALID", activation_fn=tf.nn.relu)
_layer1_1 = conv2d(_layer1_0, num_outputs=16, kernel_size=(3, 3), padding="VALID", activation_fn=tf.nn.relu)
_layer1_2 = conv2d(_layer1_1, num_outputs=16, kernel_size=(1, 1), padding="VALID", activation_fn=tf.nn.relu)
_layer1_3 = dropout(_layer1_2, keep_prob=keep_prob)
layer1 = max_pool2d(_layer1_3, (2, 2), padding="VALID")
# Layer 2: Inseption
_layer2_1 = avg_pool2d(layer1, kernel_size=(3,3), stride=1, padding="SAME")
_layer2_1 = conv2d(_layer2_1, num_outputs=32, kernel_size=(1, 1), padding="SAME", activation_fn=tf.nn.relu)
_layer2_3 = conv2d(layer1, num_outputs=32, kernel_size=(3, 3), padding="SAME", activation_fn=tf.nn.relu)
_layer2_5 = conv2d(layer1, num_outputs=32, kernel_size=(5, 5), padding="SAME", activation_fn=tf.nn.relu)
_layer2 = tf.concat([_layer2_1, _layer2_3, _layer2_5], axis=3)
_layer2 = dropout(_layer2, keep_prob=keep_prob)
layer2 = max_pool2d(_layer2, (2, 2), padding="VALID")
# Layer out: fully connected - concat the flattened previous layersoutputs
flatten_in = tf.concat([flatten(_layer1_3), flatten(layer2)], axis=-1)
flatten_1 = fully_connected(flatten_in, num_outputs=2*out_shape, activation_fn=tf.nn.relu, weights_regularizer=l2_regularizer(scale=1e-3))
flatten_1 = dropout(flatten_1, keep_prob=keep_prob)
# ouput logits
logits = fully_connected(flatten_1, num_outputs=out_shape, activation_fn=None, weights_regularizer=l2_regularizer(scale=1e-3))
layers = {"_layer1_0": _layer1_0, "_layer1_1": _layer1_1, "_layer1_2": _layer1_2, "_layer1_3": _layer1_3, "_layer2_1": _layer2_1, "_layer2_3": _layer2_3, "_layer2_5": _layer2_5, "layer_2": layer2}
return logits, layers
Training Pipeline¶
In [17]:
x = tf.placeholder(tf.float32, (None, 32, 32, 3), name="input")
y = tf.placeholder(tf.int32, (None), name="labels")
keep_prob = tf.placeholder(tf.float32,shape=(), name="keep_prob")
In [18]:
one_hot_y = tf.one_hot(y, len(set(y_train)))
rate = 1e-3
logits, network_layers = network(x, out_shape=len(set(y_train)), keep_prob=True, train=True)
#logits = network(x, out_shape=len(set(y_train)), keep_prob=keep_prob)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=one_hot_y, logits=logits)
# CE + L2 for regularization
loss_operation = tf.reduce_mean(cross_entropy + tf.losses.get_regularization_loss())
optimizer = tf.train.AdamOptimizer(learning_rate=rate)
training_operation = optimizer.minimize(loss_operation)
probs = tf.nn.softmax(logits)
Train, Validate and Test the Model¶
A validation set can be used to assess how well the model is performing. A low accuracy on the training and validation sets imply underfitting. A high accuracy on the training set but low accuracy on the validation set implies overfitting.
In [19]:
### Train your model here.
### Calculate and report the accuracy on the training and validation set.
### Once a final model architecture is selected,
### the accuracy on the test set should be calculated and reported as well.
### Feel free to use as many code cells as needed.
Model Evaluation¶
In [20]:
BATCH_SIZE = 256
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(one_hot_y, 1))
accuracy_operation = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
saver = tf.train.Saver()
def evaluate(sess, X_data, y_data, kp):
num_examples = len(X_data)
total_accuracy = 0
#sess = tf.get_default_session()
for offset in range(0, num_examples, BATCH_SIZE):
batch_x, batch_y = X_data[offset:offset+BATCH_SIZE], y_data[offset:offset+BATCH_SIZE]
accuracy = sess.run(accuracy_operation, feed_dict={x: batch_x, y: batch_y, keep_prob: kp})
total_accuracy += (accuracy * len(batch_x))
return total_accuracy / num_examples
Train the Model¶
In [21]:
if not os.path.exists("network.meta"):
EPOCHS = 10
from tqdm import tqdm
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
sess.run(tf.global_variables_initializer())
num_examples = len(X_train)
train_acc = []
valid_acc = []
for i in range(EPOCHS):
X_train, y_train = shuffle(X_train, y_train)
for offset in tqdm(range(0, num_examples, BATCH_SIZE)):
end = offset + BATCH_SIZE
batch_x, batch_y = X_train[offset:end], y_train[offset:end]
sess.run([training_operation], feed_dict={x: batch_x, y: batch_y, keep_prob: 0.5})
training_accuracy = evaluate(sess, X_train, y_train, kp=0.5)
train_acc.append(training_accuracy)
validation_accuracy = evaluate(sess, X_valid, y_valid, kp=1)
valid_acc.append(validation_accuracy)
print("EPOCH {} val acc {:.3f}".format(i+1, validation_accuracy))
saver.save(sess, './network')
print("Model saved")
with open('accuracy.pkl', 'wb') as fd:
pickle.dump({"train": train_acc, "valid": valid_acc}, fd)
In [22]:
df_results = pd.DataFrame.from_dict(pd.read_pickle("accuracy.pkl"))
fig, ax = plt.subplots(figsize=(20,20))
ax = df_results.plot(ax=ax)
ax.tick_params(axis='y', colors='g')
ax.tick_params(axis='x', colors='g')
ax.set_title("Accuracy per EPOCH", color='g', fontsize=20)
ax.set_xlabel("EPOCH", color='g', fontsize=20)
ax.set_ylabel("Accuracy", color='g', fontsize=20)
ax.grid()
plt.savefig("output_images/3_training_accuracy.png")
Final accuracy for Validation and test¶
In [23]:
with tf.Session() as sess:
saver = tf.train.import_meta_graph('./network.meta')
saver.restore(sess,tf.train.latest_checkpoint('./'))
validation_accuracy = evaluate(sess, X_valid, y_valid, kp=1)
print("val acc {:.3f}".format(validation_accuracy))
test_accuracy = evaluate(sess, X_test, y_test, kp=1)
print("test acc {:.3f}".format(test_accuracy))
P = sess.run(probs, feed_dict={x: X_test, y: y_test, keep_prob: 1})
P_web = sess.run(probs, feed_dict={x: X_web, y: y_web, keep_prob: 1})
web_accuracy = evaluate(sess, X_web, y_web, kp=1)
In [24]:
# taken from https://stackoverflow.com/questions/26678467/export-a-pandas-dataframe-as-a-table-image
def render_mpl_table(data, col_width=3.0, row_height=0.625, font_size=14,
header_color='#40466e', row_colors=['#f1f1f2', 'w'], edge_color='w',
bbox=[0, 0, 1, 1], header_columns=0,
ax=None, **kwargs):
if ax is None:
size = (np.array(data.shape[::-1]) + np.array([0, 1])) * np.array([col_width, row_height])
fig, ax = plt.subplots(figsize=size)
ax.axis('off')
mpl_table = ax.table(cellText=data.values, bbox=bbox, colLabels=data.columns, **kwargs)
mpl_table.auto_set_font_size(False)
mpl_table.set_fontsize(font_size)
for k, cell in six.iteritems(mpl_table._cells):
cell.set_edgecolor(edge_color)
if k[0] == 0 or k[1] < header_columns:
cell.set_text_props(weight='bold', color='w')
cell.set_facecolor(header_color)
else:
cell.set_facecolor(row_colors[k[0]%len(row_colors) ])
return ax
In [25]:
def show_top_5(probs, image, ax):
p = pd.DataFrame(probs, columns=["prob"])
p["name"] = list(map(lambda pid: signnames.loc[pid][0], range(len(probs))))
p = p[["name", "prob"]]
top5 = p.nlargest(5, "prob", keep="last")
top5.iloc[::-1].plot.barh(x='name', y='prob', ax=ax[0], fontsize=20)
ax[0].tick_params(axis='y', colors='g')
ax[0].tick_params(axis='x', colors='g')
ax[0].set_title("Pmax = %2.2f [%%]" % (100 * top5.prob.max()), fontsize=20)
ax[1].imshow(image)
ax[2] = render_mpl_table(top5, ax=ax[2])
return ax
n = 10
nx = 3
ny = 5
fig, ax = utils.setup_figure(x_pics=nx, y_pics=ny, w=w, h=h, n=n, sharex=False, sharey=False)
visited = []
for isample in range(ny):
j = np.random.randint(ny)
while j in visited:
j = np.random.randint(ny)
visited.append(j)
ax[isample] = show_top_5(P[j], X_test[j], ax[isample])
# plt.show()
fig.savefig('output_images/2_P5max_testset.png', bbox_inches='tight')
Step 3: Test a Model on New Images¶
To give yourself more insight into how your model is working, download at least five pictures of German traffic signs from the web and use your model to predict the traffic sign type.
You may find signnames.csv useful as it contains mappings from the class id (integer) to the actual sign name.
Load and Output the Images¶
In [26]:
### Load the images and plot them here.
### Feel free to use as many code cells as needed.
n = 10
nx = 3
ny = int(np.ceil(len(X_web_orig) / nx))
fig, ax = utils.setup_figure(x_pics=nx, y_pics=ny, w=w, h=h, n=n, sharex=False, sharey=False)
for i, orig_image in enumerate(X_web_orig):
ix = i % nx
iy = int((i - ix) / nx)
ax[iy][ix].imshow(orig_image, cmap='gray')
ax[iy][ix].set_title(signnames.loc[y_web[i], "SignName"], fontsize=20, color='g')
Predict the Sign Type for Each Image¶
In [27]:
### Run the predictions here and use the model to output the prediction for each image.
### Make sure to pre-process the images with the same pre-processing pipeline used earlier.
### Feel free to use as many code cells as needed.
with tf.Session() as sess:
saver = tf.train.import_meta_graph('./network.meta')
saver.restore(sess,tf.train.latest_checkpoint('./'))
results = pd.DataFrame(columns=["image", "label", "predicted", "probability [%]"], index=range(len(P_web)))
for iimage, index in enumerate(sess.run(tf.argmax(P_web, axis=1))):
results.loc[iimage, "image"] = os.path.split(image_list[iimage])[-1]
results.loc[iimage, "label"] = y_web[iimage]
results.loc[iimage, "predicted"] = signnames.loc[index, "SignName"]
results.loc[iimage, "probability [%]"] = 100*max(P_web[iimage])
results.to_html("output_images/3_P5max_webimages_probabilities.html")
render_mpl_table(results)
plt.savefig("output_images/3_P5max_webimages_probabilities.png")
Analyze Performance¶
In [28]:
### Calculate the accuracy for these 5 new images.
### For example, if the model predicted 1 out of 5 signs correctly, it's 20% accurate on these new images.
### Run the predictions here and use the model to output the prediction for each image.
### Make sure to pre-process the images with the same pre-processing pipeline used earlier.
### Feel free to use as many code cells as needed.
print("Web images accuracy {:.3f}".format(web_accuracy))
Output Top 5 Softmax Probabilities For Each Image Found on the Web¶
In [29]:
### Print out the top five softmax probabilities for the predictions on the German traffic sign images found on the web.
### Feel free to use as many code cells as needed.
In [30]:
n = 10
nx = 3
ny = len(y_web)
fig, ax = utils.setup_figure(x_pics=nx, y_pics=ny, w=w, h=h, n=n, sharex=False, sharey=False)
for isample in range(ny):
ax[isample] = show_top_5(P_web[isample], X_web_orig[isample], ax[isample])
fig.savefig('output_images/3_P5max_webset.png', bbox_inches='tight')
Note: Once you have completed all of the code implementations and successfully answered each question above, you may finalize your work by exporting the iPython Notebook as an HTML document. You can do this by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
Step 4 (Optional): Visualize the Neural Network's State with Test Images¶
This Section is not required to complete but acts as an additional excersise for understaning the output of a neural network's weights. While neural networks can be a great learning device they are often referred to as a black box. We can understand what the weights of a neural network look like better by plotting their feature maps. After successfully training your neural network you can see what it's feature maps look like by plotting the output of the network's weight layers in response to a test stimuli image. From these plotted feature maps, it's possible to see what characteristics of an image the network finds interesting. For a sign, maybe the inner network feature maps react with high activation to the sign's boundary outline or to the contrast in the sign's painted symbol.
Provided for you below is the function code that allows you to get the visualization output of any tensorflow weight layer you want. The inputs to the function should be a stimuli image, one used during training or a new one you provided, and then the tensorflow variable name that represents the layer's state during the training process, for instance if you wanted to see what the LeNet lab's feature maps looked like for it's second convolutional layer you could enter conv2 as the tf_activation variable.
For an example of what feature map outputs look like, check out NVIDIA's results in their paper End-to-End Deep Learning for Self-Driving Cars in the section Visualization of internal CNN State. NVIDIA was able to show that their network's inner weights had high activations to road boundary lines by comparing feature maps from an image with a clear path to one without. Try experimenting with a similar test to show that your trained network's weights are looking for interesting features, whether it's looking at differences in feature maps from images with or without a sign, or even what feature maps look like in a trained network vs a completely untrained one on the same sign image.
Your output should look something like this (above)
In [31]:
### Visualize your network's feature maps here.
### Feel free to use as many code cells as needed.
# image_input: the test image being fed into the network to produce the feature maps
# tf_activation: should be a tf variable name used during your training procedure that represents the calculated state of a specific weight layer
# activation_min/max: can be used to view the activation contrast in more detail, by default matplot sets min and max to the actual min and max values of the output
# plt_num: used to plot out multiple different weight feature map sets on the same block, just extend the plt number for each new feature map entry
sess= tf.Session()
def outputFeatureMap(image_input, tf_activation, activation_min=-1, activation_max=-1 ,plt_num=1, yplots=8):
# Here make sure to preprocess your image_input in a way your network expects
# with size, normalization, ect if needed
# image_input =
# Note: x should be the same name as your network's tensorflow data placeholder variable
# If you get an error tf_activation is not defined it may be having trouble accessing the variable from inside a function
activation = tf_activation.eval(session=sess,feed_dict={x : image_input, keep_prob: 1})
featuremaps = activation.shape[3]
plt.figure(plt_num, figsize=(15,15))
for featuremap in range(featuremaps):
#plt.subplot(int(np.ceil(featuremaps / 8)), 8, featuremap+1) # sets the number of feature maps to show on each row and column
plt.subplot(yplots, 8, featuremap+1) # sets the number of feature maps to show on each row and column
plt.title('FeatureMap ' + str(featuremap)) # displays the feature map number
if activation_min != -1 & activation_max != -1:
plt.imshow(activation[0,:,:, featuremap], interpolation="nearest", vmin =activation_min, vmax=activation_max, cmap="gray")
elif activation_max != -1:
plt.imshow(activation[0,:,:, featuremap], interpolation="nearest", vmax=activation_max, cmap="gray")
elif activation_min !=-1:
plt.imshow(activation[0,:,:, featuremap], interpolation="nearest", vmin=activation_min, cmap="gray")
else:
plt.imshow(activation[0,:,:, featuremap], interpolation="nearest", cmap="gray")
index = np.random.randint(len(X_test))
_x_test = np.expand_dims(X_test[index], 0)
_y_test = np.array([y_test[index]])
saver.restore(sess,tf.train.latest_checkpoint('./'))
In [32]:
#layers = {"_layer1_0": _layer1_0, "_layer1_1": _layer1_1, "_layer1_2": _layer1_2, "_layer1_3": _layer1_3, "_layer2_1": _layer2_1, "_layer2_3": _layer2_3, "_layer2_5": _layer2_5, "layer_2": layer2}
print("Convolutional layer _layer1_0")
outputFeatureMap(_x_test, network_layers["_layer1_0"])
plt.savefig("output_images/_layer1_0", bbox_inches='tight')
In [33]:
print("Convolutional layer _layer1_1")
outputFeatureMap(_x_test, network_layers["_layer1_1"])
plt.savefig("output_images/_layer1_1", bbox_inches='tight')
In [34]:
print("Convolutional layer _layer1_2")
outputFeatureMap(_x_test, network_layers["_layer1_2"])
plt.savefig("output_images/_layer1_2", bbox_inches='tight')
In [35]:
print("Convolutional layer _layer1_3")
outputFeatureMap(_x_test, network_layers["_layer1_3"])
plt.savefig("output_images/_layer1_3", bbox_inches='tight')
In [36]:
print("Inception layer output _layer2_1")
outputFeatureMap(_x_test, network_layers["_layer2_1"])
plt.savefig("output_images/_layer2_1", bbox_inches='tight')
In [37]:
print("Inception layer output _layer2_3")
outputFeatureMap(_x_test, network_layers["_layer2_3"])
plt.savefig("output_images/_layer2_3", bbox_inches='tight')
In [38]:
print("Inception layer output _layer2_5")
outputFeatureMap(_x_test, network_layers["_layer2_5"])
plt.savefig("output_images/_layer2_5", bbox_inches='tight')
In [39]:
print("Inception layer 2")
outputFeatureMap(_x_test, network_layers['layer_2'], plt_num=1, yplots=12)
plt.savefig("output_images/layer2", bbox_inches='tight')
In [40]:
sess.close()